[ A watcher, not a weapon ]

Watch the sky.

Research prototypeSoftware tested on recorded video and in simulation. Physical hardware validation pending.

Passive visual sensing that tells drones from aircraft and logs every sighting.

View the code See where it stands

Estimated sensing-hardware parts cost: about $113 (camera, servos, mount, power), before the computer that runs the model. Not a validated system.

SAMPLE · SIMULATED SKY
  • Detection only
  • ~$113 est. sensing parts
  • Tested in simulation and on recorded video
  • Not yet on real hardware
VARN LABS

What we've tested

DONE · SOFTWARE

Software tests

  • Detection, tracking and pan-tilt control loop, in simulation
  • Sighting logger and live dashboard, in simulation
  • Pico servo firmware, tested in software

MEASURED · RECORDED VIDEO

Recorded-video results

  • Model v2 on 43 real-world clips
  • 14.9% of drone clips labeled correctly: v2 is weak at drones
  • 7.5% of aircraft frames wrongly called a drone

PENDING · HARDWARE

Hardware tests

  • Physical node not built yet
  • No outdoor test yet
  • No comparison against real flight data (ADS-B) yet

Full results and limits ↓

01-00[ The math is backwards ]

A small drone costs a few hundred dollars. Systems built to detect one cost tens of thousands.

A small drone
$100s
Typical detection system
$10,000s
Estimated Varn sensing-hardware parts (excludes the computer)
~$113

Price comparison only, not a capability comparison. Varn is one passive camera node and has not run on real hardware yet.

Varn is a first step toward making that sensing layer cheap enough to put anywhere.

USE CASES

Early use cases I'm exploring: farms, small airports, stadiums, and power substations. First hypothesis to test: small general-aviation airfields.

02-00[ How it works ]

From sky to sighting

Five steps turn a patch of sky into a logged record. Keep scrolling and watch one happen.

  1. 01

    SEE

    A webcam on a pan-tilt mount looks at the sky. Varn is passive: it only receives light. It never jams, never transmits on aircraft or drone frequencies, and carries no payload.

    Passive · receives light only
  2. 02

    DETECT

    Every frame goes through a YOLO object-detection model. For each thing it finds, it returns a class (drone or aircraft), a confidence score and a box. Today that is model v2, which is weak at drones. Model v3 is in training.

    YOLO · ONNX model
  3. 03

    TRACK

    The tracker picks one target and stays locked on it from frame to frame. A small controller turns how far off-centre the target is into new pan and tilt angles, and a Raspberry Pi Pico drives the two servos.

    Pico · two servos
  4. 04

    LOG

    Every sighting is saved with its time, what it was, how confident the model was and which way the camera was pointing, plus a snapshot of its best moment. An hourly sky-clear check records the quiet too.

    Time · class · confidence · direction
  5. 05

    CHECK

    Planes publicly broadcast their positions (ADS-B). Varn can line its sightings up with a log of the real flights overhead, which is how real accuracy numbers will be measured. That comparison is written but has not run against a real feed yet.

    ADS-B · built, not yet run on a real feed
SAMPLE · NODE-01 · SIMULATED 01 / 05
SAMPLE: everything in this panel is a simulation, not footage.

03-00[ Reading the screen ]

Three labels, three shapes

Colour is never the only signal. Every label pairs a colour with a shape and a word, so it reads the same in greyscale and for colour-blind eyes.

  • Confirmed aircraft

    Filled circle

    The model sees an aircraft and a matching ADS-B broadcast is nearby.

  • Possible drone

    Filled square

    The model says drone. It is always only possible: a model can be wrong, and Varn says so.

  • No ADS-B match

    Hollow ring

    Looks like an aircraft, but no matching broadcast was found. It may be a plane that does not broadcast, or a misread.

The ADS-B comparison is written but has not yet been run against a real feed, so these labels are shown here as a design, not as measured results.

04-00[ Under the hood ]

The hardware and the view

SAMPLE · SENSOR VIEW · SIMULATED NODE-01
SAMPLE: a simulation of the overlay the node draws on its camera feed (brain/overlay.py). The locked target gets the bold box and a solid label; anything else gets a thin box. The camera turns to keep the lock in the centre ring.
SAMPLE · SKY PICTURE · SIMULATED N ↑
SAMPLE: made-up blips showing how sightings could be plotted around the node. Circle = aircraft, square = possible drone. Varn has no radar; this is a map of what the camera reports.
3D render of the Varn pan-tilt mount from its CAD files: a base holding the pan servo, a yoke on top, and a camera arm between the yoke's uprights.

PAN-TILT MOUNT

  • BASE

    Holds the pan servo, shaft up, and screws down with four M3 screws.

  • YOKE

    Sits on the pan horn, holds the tilt servo, and carries the M3 pivot.

  • CAMERA ARM

    Webcam cradle on the tilt horn and pivot.

Rendered live from the CAD files in this repo, with the two servos drawn as outlines, and moved through its pan and tilt axes. The physical build is next.

TEST RANGE

The test range is free: real air traffic passes overhead all day. Planes publicly broadcast their positions (ADS-B), so I can compare what the camera saw with what actually flew over and report real accuracy numbers.

05-00[ Where it stands ]

Status

  1. Pico servo firmware: smooth motion, calibration, command protocol

    DONE

  2. Detection and tracking loop, working in simulation, 243 automated tests

    DONE

  3. Pan-tilt mount designed in CAD, with STEP and STL files

    DONE

  4. Wiring diagram and bill of materials

    DONE

  5. Sighting logger (built and tested in simulation)

    DONE

  6. Live dashboard (built and tested in simulation)

    DONE

  7. Detection model v3 with drone and aircraft classes (in training)

    IN PROGRESS

  8. Run the logger and dashboard on real hardware

    NEXT

  9. Build the physical hardware

    NEXT

  10. Outdoor test

    NEXT

  11. Test against real flight data (ADS-B) and publish accuracy results

    NEXT

  12. Pi 4 + solar deployment

    NEXT

06-00[ Recorded video test ]

Results so far

The v2 model, run on 43 real-world videos of planes, military jets and drones. This tests the model on recorded video. It is not a hardware test.

False-drone rate on real aircraft
7.5%
798 of 10,657 frames were called a drone
Drone clips labeled correctly
14.9%
v2 is weak at drones. v3 is training to fix this.
Civilian planes called drones
0.8%
Military jets called drones
10.8%
Mostly distant F-35s

Model v2, 43 real-world videos. v3 (drone and aircraft classes, 960 px input) is in training. There are no v3 results yet.

07-00[ Honest limits ]

Known limits

  • RANGE · ESTIMATE

    A wide-lens webcam detects small drones only at short range, likely tens of meters (an estimate, still to be measured). Aircraft are detectable much farther.

  • CONDITIONS

    Visual sensing is weaker in darkness, fog and rain.

  • NOT REMOTE ID

    It can see drones that broadcast nothing, unlike Remote ID, but it does not replace Remote ID or RF sensors.

  • NOT BUILT YET

    Nothing has run on real hardware yet.

08-00[ About me ]

I'm Braden, and I like building things that work in the real world. I started with Sunnode, a small server that runs entirely on solar power from a panel on my fence. Then I got curious about drones, trained my first detection model, and Varn was the next step. I've run a backyard gardening business, and I shelved a software startup after learning how hard it is to sell to companies when you're young. I'm interested in how technology can make shared airspace safer, and Varn is where I'm starting.

09-00[ Roadmap ]

From one node to a network

This is Phase 1. The long-term goal is affordable, passive airspace awareness: knowing what is flying overhead, openly measured and documented.

  1. PHASE 1 · NOW

    One working node

    A camera that detects, tracks, and logs aircraft and drones, with real accuracy numbers.

  2. PHASE 2

    A network of nodes

    Several nodes working together can work out where a drone is and where it is heading.

  3. PHASE 3

    Wider awareness

    Low-cost passive sensing for places that need to know what is in their airspace, like small airfields, with published accuracy at every step.

Phase 1 is passive by design. It carries no payloads, never jams, and never transmits on aircraft or drone frequencies. It only watches.